Motivation
The simplistic answer is that a C++ module is like a header that is also a translation unit. It is like a header in that you can use it (with import, which is a new contextual keyword) to gain access to declarations from a library. Because it is a translation unit (or several for a complicated module), it is compiled separately and only once. (Recall that #include literally copies the contents of a file into the translation unit that contains the directive.) This combination yields a number of advantages:
- Isolation: because a module unit is a separate translation unit, it has its own set of macros and
using declarations/directives that neither affect nor are affected by those in the importing translation unit or any other module. This prevents collisions between an identifier #defined in one header and used in another. While use of using still should be judicious, it is not intrinsically harmful to write even using namespace at namespace scope in a module interface.
- Interface control: because a module unit can declare entities with internal linkage (with
static or namespace {}), with export (the keyword reserved for purposes like these since C++98), or with neither, it can restrict how much of its contents are available to clients. This replaces the namespace detail idiom which can conflict between headers (that use it in the same containing namespace).
- Deduplication: because in many cases it is no longer necessary to provide a declaration in a header file and a definition in a separate source file, redundancy and the associated opportunity for divergence are reduced.
- One Definition Rule violation avoidance: the ODR exists solely because of the need to define certain entities (types, inline functions/variables, and templates) in every translation unit that uses them. A module can define an entity just once and nonetheless provide that definition to clients. Also, existing headers that already violate the ODR via internal-linkage declarations stop being ill-formed, no diagnostic required, when they are converted into modules.
- Non-local variable initialization order: because
import establishes a dependency order among translation units that contain (unique) variable definitions, there is an obvious order in which to initialize non-local variables with static storage duration. C++17 supplied inline variables with a controllable initialization order; modules extend that to normal variables (and do not need inline variables at all).
- Module-private declarations: entities declared in a module that neither are exported nor have internal linkage are usable (by name) by any translation unit in the module, providing a useful middle ground between the preexisting choices of
static or not. While it remains to be seen what exactly implementations will do with these, they correspond closely to the notion of “hidden” (or “not exported”) symbols in a dynamic object, providing a potential language recognition of this practical dynamic linking optimization.
- ABI stability: the rules for
inline (whose ODR-compatibility purpose is not relevant in a module) have been adjusted to support (but not require!) an implementation strategy where non-inline functions can serve as an ABI boundary for shared library upgrades.
- Compilation speed: because the contents of a module do not need to be reparsed as part of every translation unit that uses them, in many cases compilation proceeds much faster.
- Tooling: the “structural declarations” involving
import and module have restrictions on their use to make them readily and efficiently detectable by tools that need to understand the dependency graph of a project. The restrictions also allow most if not all existing uses of those common words as identifiers.
Approach
Because a name declared in a module must be found in a client, a significant new kind of name lookup is required that works across translation units; getting correct rules for argument-dependent lookup and template instantiation was a significant part of what made this proposal take over a decade to standardize. The simple rule is that (aside from being incompatible with internal linkage for obvious reasons) export affects only name lookup; any entity available via (e.g.) decltype or a template parameter has exactly the same behavior regardless of whether it is exported.
Because a module must be able to provide types, inline functions, and templates to its clients in a way that allows their contents to be used, typically a compiler generates an artifact when processing a module (sometimes called a Compiled Module Interface) that contains the detailed information needed by the clients. The CMI is similar to a pre-compiled header, but does not have the restrictions that the same headers must be included, in the same order, in every relevant translation unit. It is also similar to the behavior of Fortran modules, although there is no analog to their feature of importing only particular names from a module.
Because the compiler must be able to find the CMI based on import foo; (and find source files based on import :partition;), it must know some mapping from “foo” to the (CMI) file name. Clang has established the term “module map” for this concept; in general, it remains to be seen just how to handle situations like implicit directory structures or module (or partition) names that don’t match source file names.
Non-features
Like other “binary header” technologies, modules should not be taken to be a distribution mechanism (as much as those of a secretive bent might want to avoid providing headers and all the definitions of any contained templates). Nor are they “header-only” in the traditional sense, although a compiler could regenerate the CMI for each project using a module.
While in many other languages (e.g., Python), modules are units not only of compilation but also of naming, C++ modules are not namespaces. C++ already has namespaces, and modules change nothing about their usage and behavior (partly for backward compatibility). It is to be expected, however, that module names will often align with namespace names, especially for libraries with well-known namespace names that would be confusing as the name of any other module. (A nested::name may be rendered as a module name nested.name, since . and not :: is allowed there; a . has no significance in C++20 except as a convention.)
Modules also do not obsolete the pImpl idiom or prevent the fragile base class problem. If a class is complete for a client, then changing that class still requires recompiling the client in general.
Finally, modules do not provide a mechanism to provide the macros that are an important part of the interface of some libraries; it is possible to provide a wrapper header that looks like
// wants_macros.hpp
import wants.macros;
#define INTERFACE_MACRO(x) (wants::f(x),wants::g(x))
(You don't even need #include guards unless there might be other definitions of the same macro.)
Multi-file modules
A module has a single primary interface unit that contains export module A;: this is the translation unit processed by the compiler to produce the data needed by clients. It may recruit additional interface partitions that contain export module A:sub1;; these are separate translation units but are included in the one CMI for the module. It is also possible to have implementation partitions (module A:impl1;) that can be imported by the interface without providing their contents to clients of the overall module. (Some implementations may leak those contents to clients anyway for technical reasons, but this never affects name lookup.)
Finally, (non-partition) module implementation units (with simply module A;) provide nothing at all to clients, but can define entities declared in the module interface (which they implicitly import). All translation units of a module can use anything declared in another part of the same module that they import so long as it does not have internal linkage (in other words, they ignore export).
As a special case, a single-file module can contain a module :private; declaration that effectively packages an implementation unit with the interface; this is called a private module fragment. In particular, it can be used to define a class while leaving it incomplete in a client (which provides binary compatibility but will not prevent recompilation with typical build tools).
Upgrading
Converting a header-based library to a module is neither a trivial nor a monumental task. The required boilerplate is very minor (two lines in many cases), and it is possible to put export {} around relatively large sections of a file (although there are unfortunate limitations: no static_assert declarations or deduction guides may be enclosed). Generally, a namespace detail {} can either be converted to namespace {} or simply left unexported; in the latter case, its contents may often be moved to the containing namespace. Class members need to be explicitly marked inline if it is desired that even ABI-conservative implementations inline calls to them from other translation units.
Of course, not all libraries can be upgraded instantaneously; backward comptibility has always been one of C++’s emphases, and there are two separate mechanisms to allow module-based libraries to depend on header-based libraries (based on those supplied by initial experimental implementations). (In the other direction, a header can simply use import like anything else even if it is used by a module in either fashion.)
As in the Modules Technical Specification, a global module fragment may appear at the beginning of a module unit (introduced by a bare module;) that contains only preprocessor directives: in particular, #includes for the headers on which a module depends. It is possible in most cases to instantiate a template defined in a module that uses declarations from a header it includes because those declarations are incorporated into the CMI.
There is also the option to import a “modular” (or importable) header (import "foo.hpp";): what is imported is a synthesized header unit that acts like a module except that it exports everything it declares—even things with internal linkage (which may (still!) produce ODR violations if used outside the header) and macros. (It is an error to use a macro given different values by different imported header units; command-line macros (-D) aren't considered for that.)
